JudgeMoE: Distributional Aggregation for LLM-as-a-Judge

JudgeMoE, a lightweight aggregator that assigns example-specific weights to cached LLM judge score distributions before fusing them, improved mean Spearman correlation over uniform log pooling by +0.079 on the original 10-cell benchmark, according to the arXiv paper 2610.07109v1. Applied to six additional cells, the same configuration delivered a +0.0393 mean gain over the strongest local single judge across 16 cells, with positive differences in 12/16 cells and a one-sided Wilcoxon signed-rank p=0.0091. The authors' protocol study also found score-range choice unstable across judge-dataset settings and soft scoring usually outperforming hard decoding, with validation-based analyses showing the preferred aggregation method depends on the task and judge pool.

arXiv:2610.07109v1 Announce Type: new Abstract: When an LLM judge scores an output, its score distribution retains uncertainty and disagreement information that is lost after scalar compression. We introduce JudgeMoE, a lightweight aggregator that assigns example-specific weights to cached judge score distributions and fuses them before computing a final score. A protocol study shows that score-range choice is unstable across judge--dataset settings and that soft scoring usually outperforms hard decoding. On the original 10-cell benchmark, JudgeMoE improves mean Spearman over uniform log pooling by $+0.079$. Applying the same configuration to six additional cells yields a $+0.0393$ mean gain over the strongest local single judge across 16 cells, with positive differences in 12/16 cells and a one-sided Wilcoxon signed-rank $p=0.0091$. Validation-based analyses further show that the preferred aggregation method depends on the task and judge pool.